Fix: ObjectBoxVectorIndex hardcoded 768 dimensions breaks TFLite/ONNX

[iberi22]

Problem

ObjectBoxVectorIndex in lib/src/vector_index_objectbox.dart has hardcoded dimension of 768:

@HnswIndex(dimensions: 768, distanceType: VectorDistanceType.cosine)
@Property(type: PropertyType.floatVector)
List<double>? vector;

But TFLite embedding models (e.g., universal-sentence-encoder-lite) produce 384-dimensional vectors. ONNX models like all-MiniLM-L6-v2 also produce 384 dimensions.

This means switching from Gemini (768-dim) to TFLite/ONNX will crash at runtime because ObjectBox expects 768-dim vectors.

Root Cause

The HNSW index was designed only for Gemini's text-embedding-004 (768-dim). When using on-device models with different dimensions, the index fails.

Proposed Solution

1. Make dimension configurable via entity factory:

   • Create separate ObxVectorDoc entities per dimension
   • Or use a single entity with dynamic dimension validation

2. Update VectorIndex interface to include dimension property

3. Update memory_graph.dart to use backend's actual dimension during initialization

Impact

Fixes compatibility with:

• TFLite universal-sentence-encoder-lite (384-dim)
• ONNX all-MiniLM-L6-v2 (384-dim)
• Any future embedding model with different dimensions

Files Affected

• lib/src/vector_index_objectbox.dart - entity definition
• lib/src/memory_graph.dart - uses hardcoded dimension in initialize()

[google-labs-jules]

Jules is on it. When finished, you will see another comment and be able to review a PR.

[iberi22]

@jules-ai después de merge, necesito que hagas optimization pass en todos los componentes: weight, speed, stability. Ejecutar tests completos del paquete. Empezar con #41 primero.

[google-labs-jules]

Ready for a review! A PR has been created.